A new study published in Computational Urban Science demonstrates that large language models can effectively extract geographically grounded sentiment and urban landmarks from unstructured guest reviews. The research, conducted by Haozheng Zu, Richard Grant, and Shouraseni Sen Roy, applies a computational and theoretical framework to investigate digitally mediated urban experiences. By analyzing a large-scale dataset of Airbnb user-generated content, the authors reconceptualize urban legibility as a fluid, hybrid outcome. This outcome is produced by the interaction of material form, platform infrastructures, and aggregated user narratives, rather than as a static property of the built environment.
The study utilizes 365,365 Airbnb reviews collected between 2010 and 2024 from two contrasting geographic contexts: Broward County, Florida, and Cape Town, South Africa. These locations were selected to explore how visitors interpret and navigate diverse urban environments that are globally connected yet locally distinctive. The dataset includes geographic coordinates, accommodation types, review scores, prices, rental dates, and host and reviewer names. The total number of unique hosts was 67,295 in Cape Town and 50,147 in Broward County, providing a robust basis for comparative analysis.
To process the multilingual nature of the data, the researchers employed OpenAI’s GPT-4o model to translate reviews written in over 15 languages into American English. This step was crucial for ensuring analytical reliability across the diverse linguistic landscape of the two study sites. The analysis combined spatial clustering, temporal trend analysis, and sentiment mapping to identify evolving hotspots. The goal was to examine how digitally amplified reputations attach affective meanings to specific neighborhoods, thereby reshaping urban spatial imaginaries.
The spatial analysis utilized Emerging Hot Spot Analysis, a technique that creates a space-time cube to evaluate the spatial clustering of data over time. This method involves aggregating points with geographic coordinates and a time coordinate to monitor trajectories and their geographical distribution. The analysis identified three distinct hotspot categories: consecutive, new, and sporadic. Consecutive hotspots emerged as statistically significant in the final time steps and maintained that significance for at least two consecutive intervals, indicating persistent areas of high activity.
In Broward County, significant clusters were located along the coast, with a noticeable spillover into neighboring inland areas. The expansion of new hotspots was centered on key tourist attractions, including Fort Lauderdale Airport, Port Everglades, and entertainment complexes such as the Hard Rock Casino in Hollywood. In Cape Town, the distribution of Airbnb locations was influenced by proximity to major attractions like Table Mountain and the coast, as well as factors such as safety concerns and power outages. Two additional regional clusters were prominent in Stellenbosch and Somerset West, driven by the presence of wine estates and natural hiking trails.
For sentiment analysis, the study employed NLP Town’s pre-trained, multilingual Bidirectional Encoder Representations from Transformers-based model. This model was fine-tuned on a vast multilingual corpus of product and service reviews to predict ratings, allowing for the extraction of subjective opinions. The model’s output was mapped to a normalized sentiment score between 0 and 1, enabling the calculation of average sentiment scores for each week from June 2017 to June 2023. This longitudinal analysis facilitated the examination of temporal trends in guest satisfaction and the identification of significant shifts or patterns.
The findings revealed that both locations experienced a substantial drop in visitors due to the COVID-19 pandemic during 2020 and 2021. However, the recovery patterns differed due to varying lockdown measures. Florida saw a rise in domestic travel, while South Africa’s hard lockdowns recalibrated tourism around semigration to the Western Cape. The subsequent return to near-baseline variance underscores the market resilience of both locations, primarily driven by a significant rise in revenge travel. Sentiment scores in Florida trended more positively toward the end of 2020, while Cape Town exhibited a higher aggregate sentiment of 0.72 compared to Broward County’s 0.66.
The research also highlighted the role of platform-mediated legibility in shaping visitor perceptions. Aggregated guest reviews function as digitally mediated nodes that restructure urban legibility, evolving Kevin Lynch’s traditionally static concept of imageability into a dynamic, platform-driven paradigm. In Broward County, high-density affective mentions elevated locations like Lauderdale-by-the-Sea and Las Olas Oceanside Park from transient points of interest to primary landmarks. In Cape Town, the geospatial clustering of markets such as the Oranjezicht City Farm Market generated a digital overlay that modulated tourist mobility.
Linguistic landscapes also played a significant role in urban legibility. Broward County’s prominence of English, Spanish, French, and Portuguese reviews reflects a demographic profile shaped by sustained migration from the Americas and the Caribbean. In contrast, Cape Town’s linguistic profile, which includes German, French, Dutch, and Portuguese, is shaped by longstanding historical relationships and targeted strategic marketing. These linguistic features function as social nodes, allowing visitors to develop coherent mental maps and engage effectively with the urban environment.
The study concludes that platform placemaking reorders urban salience rather than displacing the physical city. While material features such as streetscapes and physiography remain foundational, their prominence is increasingly filtered through recommendation systems and feedback loops. In this hybrid environment, imageability is distributed and dynamic, with urban elements gaining significance through social layers of repetition, affective intensity, and algorithmic ranking. The findings underscore the adaptive resilience of digital tourism markets in co-producing the contemporary city.
However, the authors acknowledge several limitations in their study. The findings have limited generalizability due to reliance on Inside Airbnb data and random sampling that may overlook local spatial differences. Computational limitations, such as potential gender misclassification and missed cultural nuances in sentiment analysis, also affect the results. Furthermore, although GPT-4o performs reliably across languages, cultural nuances may be imperfectly rendered in translation, potentially affecting sentiment interpretation and keyword identification.
Despite these constraints, the study provides valuable insights into the intersection of digital placemaking and Lynchian imageability. It demonstrates that large language models can accurately identify urban physiographic landmarks, nodes, sentiment expressions, and symbolic meanings in reviews. These insights equip urban planners to align tourism infrastructure with user-generated priorities while remaining vigilant to the fragility of algorithm-curated legibility under socio-political stress. The research offers a data-driven roadmap for policymakers and urban stakeholders to synchronize infrastructure investments with emerging guest spatial preferences.
The study’s theoretical framework revisits Kevin Lynch’s 1960 concept of “imageability,” originally defined as the quality of a physical object that evokes a strong image in an observer. The authors argue that in the contemporary digital environment, this legibility is no longer a static property of the built environment but a fluid outcome shaped by the interaction of material form, platform infrastructures, and aggregated user narratives. Algorithmic mediation now constitutes an interdependent layer of urban legibility, reshaping how paths, nodes, districts, and landmarks are perceived and circulated. This perspective suggests that city legibility emerges as a synthesis of physical space and the platform-driven attention economy, reconfiguring the mental map into a landscape that integrates tangible structures with transient digital data.
Methodologically, the research combined geostatistical methods with deep learning to analyze urban space dynamics. The authors utilized OpenAI’s GPT-4o to translate reviews into American English, addressing the multilingual nature of the dataset. For sentiment analysis, they employed NLP Town’s pre-trained, multilingual Bidirectional Encoder Representations from Transformers-based model, which was fine-tuned on a vast multilingual corpus of product and service reviews. This model predicts 1- to 5-star ratings, allowing for the extraction of subjective opinions and sentiments. The output was mapped to a normalized sentiment score between 0 and 1, facilitating the calculation of average sentiment scores for each week from June 2017 to June 2023. This longitudinal approach enabled the examination of temporal trends in guest satisfaction and the identification of significant shifts or patterns in the data.
The spatial analysis revealed that Airbnb locations in both study sites show a high concentration near coastal areas and major tourist attractions.
The findings indicate that platform placemaking reorders urban salience rather than displacing the physical city. The study concludes that algorithmic mediation constitutes an interdependent layer of urban legibility, reconceptualizing Lynchian theory for the hybrid digital-physical co-produced city.
The authors acknowledge several limitations in their study, noting that the findings have limited generalizability due to reliance on Inside Airbnb data and random sampling that may overlook local spatial differences. The two-case study design further limits transferability to other urban tourism contexts, suggesting that specific urban recommendations should be interpreted with caution given the diversity of regulatory and platform conditions.
Subject of Research: Room with a (Re)View: deep learning employing OpenAI, spatial and temporal analysis at the neighborhood level
Article Title: Room with a (Re)View: deep learning employing OpenAI, spatial and temporal analysis at the neighborhood level
Article References: Zu, H., Grant, R., & Roy, S. S. (2026). Room with a (Re)View: deep learning employing OpenAI, spatial and temporal analysis at the neighborhood level. Computational Urban Science, 6(1), Article 39. https://doi.org/10.1007/s43762-026-00274-5
Image Credits: AI Generated
DOI: 10.1007/s43762-026-00274-5
Keywords: Room, View, deep, learning, employing, OpenAI, spatial, temporal, analysis, neighborhood, level, scientific research
News Source: Blake Davidson. (November 7, 2026). Deep Learning Maps Airbnb Reviews to Urban Landmarks. Scienmag.



